Automated Plant Cell Morphology through Volumetric Image Segmentation and Placenta Pathology Analysis from Photos
Restricted (Penn State Only)
- Author:
- Davaasuren, Dolzodmaa
- Graduate Program:
- Informatics
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- February 19, 2024
- Committee Members:
- James Wang, Chair & Dissertation Advisor
Sharon Huang, Major Field Member
Alison Gernand, Outside Unit & Field Member
Justin Silverman, Major Field Member
Dongwon Lee, Program Head/Chair
Charles Anderson, Outside Field Member - Keywords:
- Segmentation
Deep Learning
Machine Learning
Medical Imaging
Data Science
Placenta
Plant Cell
Cell Morphology - Abstract:
- In recent decades, biological imaging has seen remarkable advancements driven by the evolu- tion of image processing technologies and computer vision algorithms. These developments have sparked significant interest in creating data-driven solutions with practical applica- tions. However, the analysis of biological images is inherently complex, demanding con- siderable time, effort, resources, and specialized training. The introduction of automated image understanding tools promises to revolutionize biological image analysis by enhancing performance and significantly reducing the workload associated with diagnosis, treatment, and cellular analysis for medical professionals and biologists. Within this paradigm, in the first part, we address the need for volumetric measurements of stomatal guard cells and their surrounding pavement cells when only a handful of images are available. In plants, stomatal guard cells function as dynamic gatekeepers of CO2 and water flux. Despite the central functions of these cells in transpiration and photosynthesis, our knowledge of their cell walls’ mechanics is limited. To assess the wall mechanics of the guard cells, we need to capture the whole volume accurately with an automated segmentation algorithm. Automating the 3D segmentation of stomatal guard cells and other confocal mi- croscopy data is extremely challenging due to hardware limitations, hard-to-localize regions, and limited optical resolution. We present two networks to address this problem. The first model, 3D-CellNet, presents a memory-efficient, attention-based, one-stage seg- mentation neural network for 3D images of stomatal guard cells. Our model is trained end-to-end and achieves expert-level accuracy while leveraging only eight human-labeled volume images. As a proof-of-concept, we applied our model to 3D confocal data from a cell ablation experi- ment that tests the “polar stiffening” model of stomatal biomechanics. The resulting data allow us to refine this polar stiffening model. Collecting the whole volumetric image itself is a time-consuming, challenging, and painstaking process, not to mention annotating 3D cells in the image. To make use of all the unlabeled samples collected by the biologists, we make an extension in the second model, 3D-CellNext, which presents an even more efficient solution to the data scarcity issue by leveraging self-supervised learning and local, global connectivity-based hybrid encoder to take full advantage of the small dataset collected by biologists. Our model will allow biologists to rapidly test cell mechanics and dynamics and help them identify plants that more efficiently use water, a major limiting factor in global agricultural production and an area of critical concern during climate change. In the latter part of the dissertation, we aim to tackle the problem of automatic assess- ment and examination of the human placenta from photos. Automated placenta analysis is essential for evaluating the health risks of the mother and baby after delivery. However, only about 20% of placentas in the U.S. are assessed by pathology exams [1], and placental data is often missed in pregnancy research because of the additional time, cost, and ex- pertise needed in handling and ana- lyzing them. Placentas are examined by pathologists only when necessary and when resources are available. The current health care system relies on placental pathology exams, which are costly, multi-phase, lengthy processes and require high-level training of a pathologist. A computer-based, fully automated tool that can provide an immediate and comprehensive placental assessment at the time of delivery would poten- tially improve health care and radically improve medical knowledge. Collaborating with the Northwestern University Hospital, we curated a first-of-a-kind placenta dataset with pla- centa photos, pathological diagnosis, and manually traced segmentation maps. Based on the segmentation work completed by our group, we propose automatic placental pathological indicator prediction models. We hope such prediction will aid pathologists or other clients in their decision-making process and ultimately accelerate the overall process of placenta analysis.